OneCap’s Agentic AI Surfaces ₹500+ Crore In Finance Errors

OneCap, a Bengaluru-based AI platform serving enterprise finance teams, says it surfaced over ₹500+ crore in financial discrepancies inside its first 150 days, processing more than ₹17,000 crore in transactions for customers including Malabar Gold Group, HomeLane, and Renata Precision Components. The findings came wrapped in a larger claim: up to 95% of the manual reconciliation effort for those customers is now gone.

Co-founder and CEO Sandeep Nambiar, in an interview with CXOToday, frames the catch as the validation of a thesis his company has run since incorporation in 2025. In his framing, agentic AI in finance works as a visibility layer over the work the ledger was never designed to do. The ₹500+ crore figure quantifies the size of what manual reconciliation was leaving on the table.

The 150-Day Scoreboard

OneCap’s first 150 days are the most concrete data point yet for what agentic AI does when it sits across a real enterprise finance stack. The company says it processed over ₹17,000 crore in transactions for customers including Malabar Gold Group, HomeLane, Renata Precision Components, Me N Moms, and VBazaar. Out of that volume, it surfaced more than ₹500+ crore in financial discrepancies for those finance teams to act on. Reconciliation effort, by the company’s own count, fell by up to 95%.

Sandeep Nambiar, OneCap’s co-founder and CEO, sits at the centre of the claim. He founded the company with CTO Gururaj Laxmayya after senior product roles at Perfios and Open, and registered OneCap Technologies in Karnataka in 2025 (Corporate Identity Number U62099KA2025PTC197318). The company describes itself as an AI-native financial integrity layer that sits across existing ERPs, payment gateways, and tax systems without replacing them.

Two numbers in the interview make the catch feel less like a sales pitch and more like an industry diagnosis. On the accounts-payable side alone, OneCap says it has been surfacing financial discrepancies averaging around 3.5% of payable transaction value over the past three months, each finding tagged with an explanation of how the agent reached its conclusion. Across the same customer base, the company has encountered 20-plus different ERPs, many of them local or in-house, and thousands of distinct ledger formats. The discrepancies and the system diversity describe the same thing from two angles: enterprise finance, in this dataset, leaks in places humans were never looking.

• ₹17,000 crore in transactions processed across enterprise customers in 150 days
• ₹500+ crore in financial discrepancies surfaced for follow-up
• 95% reconciliation effort removed for those customers
• 3.5% average discrepancy rate on accounts payable over the past three months
• 20+ distinct ERP systems seen across the customer base

Why Deterministic Automation Stops at the Exception

Traditional finance automation runs on rules. An invoice arrives, the system checks that the supplier matches, the amount lines up, the tax fields look right, and the payment is queued. The moment any of those checks fails, the workflow stops and a human takes over. Nambiar’s framing is blunt: deterministic automation has no judgment and no ability to handle ambiguity, which is exactly what real finance operations are full of. Rules engines were designed for clean data and break the moment reality deviates from the rule they were written for.

Agentic AI works differently. The agent reasons across context, handles exceptions, and chains together multiple steps toward an outcome instead of executing one hard-coded instruction. In Nambiar’s description, the agent behaves less like a script and more like an analyst who can pull the supporting documents, account for the deductions, and document the reasoning behind each step.

The example Nambiar uses is a textbook reconciliation mess. A customer pays three invoices in a single lump sum, deducts a TDS amount, applies a credit note from last quarter, and references none of it cleanly. A rules engine flags the whole thing as an exception and hands it to a person. An agent investigates, pulls the related invoices, accounts for the TDS, locates the credit note, nets it off against the relevant invoice, excludes contra entries, and reconciles the payment. The agent then writes up what it did and why. A senior analyst, doing the same work, might take an afternoon.

The shift Nambiar names is from automating keystrokes to automating judgment. Humans still review the output; the loop just runs in seconds rather than days.

The Three Layers Manual Finance Quietly Carries

Nambiar splits enterprise finance into three layers stacked from the foundation up. Most of the work, and most of the leaks, happen in the bottom two.

The first layer is transaction sanctity. A finance team checks that invoices are valid, that payments match them, that TDS has been deducted at the right rate, and that credit notes have been applied to the right invoices. Each of those checks is mechanical on its own, but the permutations multiply quickly. A single supplier relationship can produce dozens of moving parts per month, and every part has to reconcile to a payment that may arrive days or weeks later through a different channel.

Above that sits the compliance layer. GST, e-invoicing, TDS, and frequent regulatory changes mean every transaction carries reporting obligations that have to be validated, not just recorded. India’s tax regime adds its own density here, with overlapping rules that change often enough that hand-keyed compliance work carries real risk for the finance team doing it.

Only on top of both sits the third layer: decisioning and credit. Working-capital choices, supplier financing, and short-term lending decisions all depend on the numbers underneath being reliable.

In most organisations, the bottom two layers are still manual. Nambiar’s number for the time cost is the one that lands: finance teams spend well over 80% of their hours on foundational verification and compliance work. A working-capital framework built on top of that base inherits every error the base has not yet caught. Fixing the foundation, before adding another analytics dashboard on top, is the move he argues actually changes the answer.

India Adds Its Own Pressure

India’s payments stack has scaled faster than its back offices. UPI alone now settles billions of transactions a month, and the surface area of a single sale now spans cards, net banking, wallets, BNPL, and e-commerce and quick-commerce marketplaces, each with its own data format and settlement timing. The front end of Indian commerce has been digitised and made instant. The finance function behind it has not. India’s 170 BFSI GCCs have grown into the country’s AI finance muscle over the last five years, with 60 opened in that window owning risk, fraud, and AI work, but the long tail of mid-market enterprise finance still runs on hand-stitched spreadsheets. The build-out is uneven: a thin top layer of mature capability and a thick base of work that has barely moved.

Nambiar lays out four pressures the Indian context puts on a finance team:

• Transaction volume has exploded while the unit value has shrunk, so finance teams reconcile millions of small transactions across many payment modes and intermediaries.
• The proliferation of channels, including UPI, cards, net banking, wallets, BNPL, e-com and quick-com marketplaces, means a single sale leaves fragmented traces across several systems.
• India’s compliance environment adds real density through GST, e-invoicing, TDS, and frequent regulatory change.
• Settlement and reconciliation cycles haven’t kept pace with real-time payments, so the gap between money moving and books closing has widened.

The net effect is a widening gap. Money moves in seconds; the books catch up days or weeks later, often with errors that compound. That gap is where the leakage accumulates, and it’s the slice OneCap’s agents are now scanning in full.

Interoperability Without the Migration

The traditional answer to system fragmentation was rip and replace, or a long programme of brittle point-to-point integrations that broke every time an ERP vendor pushed an upgrade. Both approaches assume the underlying systems need to be consolidated before anything new can be built on top. The Model Context Protocol open standard from Anthropic, and similar emerging protocols, invert that assumption. AI agents can sit across disconnected systems, read from each, and operate inside existing workflows as an integrity layer without forcing a costly migration.

For OneCap, MCP-class interoperability is what lets one agent layer span 20-plus ERP flavours and thousands of ledger formats without writing a custom connector for each. The agent reads from whichever system holds the source of truth, reconciles it against whichever other system holds the matching record, and pushes the result back into the workflow the finance team already runs. Interoperability stops being a precondition the enterprise has to solve before it gets value, and becomes something the AI bridges while the underlying systems stay as they are. For most enterprises, that is a far more realistic path than another multi-year platform overhaul, and a closer look at India’s 170 BFSI GCCs owning AI finance work shows where the country’s enterprise AI capacity is actually concentrating.

What Auditors Will Actually Accept

Nambiar calls traceability the thing that determines whether AI gets adopted in finance at all. Finance is not a domain where a plausible-looking answer is good enough. Every output has to withstand scrutiny from an auditor or regulator months after the agent produced it.

The combination of explainable outputs and tightly governed, fully logged operations is what turns AI in finance from a leap of faith into something a controller, an auditor, and a regulator can each independently verify.

The 3.5% figure on accounts payable is the example Nambiar uses to make that point concrete. A discrepancy rate that large, surfaced without explanation, would terrify a CFO. The same rate, surfaced with a clear chain of evidence showing which fields the agent compared, which rules it applied, and which exceptions it resolved, becomes an actionable queue a finance leader can work through and an auditor can verify. Explainability is not a feature the platform adds later; it is the property that lets the platform operate inside a regulated function at all.

The operating model matters as much as the output. Nambiar says actions should run in isolated, strictly permissioned environments, with every step logged and auditable in real time. The company was registered with that posture built in, not bolted on, and Nambiar argues that posture is the difference between an AI deployment an auditor signs off on and one that stalls in the review queue.

Five-Year Forecast From the Operator

Asked which finance functions will be AI-assisted within five years, Nambiar picks the finance controller function as the clearest candidate. He frames it as AI-enabled with a human in the loop rather than fully autonomous: reconciliation, transaction validation, compliance checks, and exception handling are well-suited to agentic AI, while the controller retains oversight and final judgment. The work the controller spends hours on now gets compressed into seconds; the work the controller actually decides on stays with the controller.

Beyond the controller’s remit, Nambiar extends the same pattern to payments operations and to contextual credit and working-capital decisions. Once transactional data is continuously verified, credit and working-capital choices can run against real activity at the moment it is needed, instead of after the close.

The path he describes is sequential. Data integrity first, workflow automation next, embedded credit last, with each layer earning the right to the next by producing the audit trail the previous one demanded. His practical advice to controllers is to start now: run focused proofs of concept, operationalise the ones that show clear return on investment, and build the muscle for knowing where AI is reliable and where human judgment is essential. The teams that build that muscle early, he argues, will hold a real advantage over those who wait.

Frequently Asked Questions

What is agentic AI in finance?

Agentic AI in finance refers to AI systems that reason across context, handle exceptions, and chain multiple steps together to complete a task, rather than executing a single hard-coded instruction. In reconciliation and compliance work, agentic systems can investigate ambiguous matches, pull supporting documents, and document their reasoning, while a human reviews the output.

How is agentic AI different from traditional finance automation?

Traditional automation runs on fixed rules and breaks the moment a real-world transaction deviates from the rule it was written for. Agentic AI works through ambiguity by reasoning over context, handling exceptions, and combining multiple steps toward an outcome. The shift, as OneCap’s Sandeep Nambiar puts it, is from automating keystrokes to automating judgment, with a human reviewing the result.

What has OneCap actually processed?

According to the company’s co-founder and CEO Sandeep Nambiar, OneCap processed over ₹17,000 crore in transactions across enterprise customers including Malabar Gold Group, HomeLane, Renata Precision Components, Me N Moms, and VBazaar in its first 150 days. The company says it surfaced more than ₹500+ crore in financial discrepancies for those customers to act on and removed up to 95% of reconciliation effort for those teams.

What is the Model Context Protocol and why does it matter here?

The Model Context Protocol is an open standard that lets AI applications connect securely to external data sources and tools. For enterprise finance, the relevance is that it lets an agent layer sit across disconnected ERPs, payment gateways, and tax systems without requiring a multi-year consolidation programme, which is the practical prerequisite for any AI that has to span the messy system landscapes OneCap describes.

OneCap’s AI-native financial integrity layer is documented on the company site, and the full OneCap interview on agentic AI in CXOToday carries every figure cited here in Nambiar’s own words.